def build(self):
"""Build the trainer by assembling the necessary components."""
super().build()
self.optimizer = Optimizer()(model=self.model)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
self.metric_name = self.config.metric.type
# Some trainer has different train batch size from valid batch
self.train_metrics = None
self.valid_metrics = self._init_metrics()
def model_fn(self, features, labels, mode):
"""Define cars model_fn used by TensorFlow Estimator."""
logging.info('Cars model function action')
self.trainer.loss = Loss()()
train_op = None
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.compat.v1.train.get_global_step()
epoch = tf.cast(global_step, tf.float32) / tf.cast(
len(self.trainer.train_loader), tf.float32)
self.trainer.optimizer = Optimizer()(
distributed=self.trainer.distributed)
self.trainer.lr_scheduler = LrScheduler()(self.trainer.optimizer)
self.trainer.lr_scheduler.step(epoch)
self.trainer.model.training = True
alphas = tf.convert_to_tensor(self.alphas)
for j in range(self.alg_policy.num_individual_per_iter):
i = np.random.randint(0, self.alg_policy.num_individual, 1)[0]
if self.epoch < self.alg_policy.warmup:
alpha = tf.convert_to_tensor(
self.search_alg.random_sample_path())
else:
alpha = alphas[i]
logits = self.trainer.model(features, alpha=alpha)
logits = tf.cast(logits, tf.float32)
loss = self.trainer.loss(logits=logits, labels=labels)
loss = self.trainer.optimizer.regularize_loss(loss)
grads, vars = zip(
*self.trainer.optimizer.compute_gradients(loss))
if j == 0:
accum_grads = [
tf.Variable(tf.zeros_like(grad), trainable=False)
for grad in grads
]
accum_grads = [
accum_grads[k] + grads[k] for k in range(len(grads))
]
if self.epoch < self.alg_policy.warmup:
break
clipped_grads, _ = tf.clip_by_global_norm(
accum_grads, self.trainer.config.grad_clip)
minimize_op = self.trainer.optimizer.apply_gradients(
list(zip(clipped_grads, vars)), global_step)
update_ops = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)
train_op = tf.group(minimize_op, update_ops)
eval_metric_ops = None
if mode == tf.estimator.ModeKeys.EVAL:
alpha = tf.convert_to_tensor(self.trainer.valid_alpha)
self.trainer.model.training = False
logits = self.trainer.model(features, alpha=alpha)
logits = tf.cast(logits, tf.float32)
loss = self.trainer.loss(logits=logits, labels=labels)
eval_metric_ops = self.trainer.valid_metrics(logits, labels)
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
def _build_multigpu_train_op(self, num_gpus):
with self.graph.as_default(), tf.device('/gpu:0'):
tower_grads = []
self.inputs = []
self.labels = []
opt = Optimizer()()
for i in range(num_gpus):
with tf.device('/gpu:%d' % i):
with tf.name_scope('tower_%d' % i) as scope:
# tf.get_variable_scope().reuse_variables()
inputs = self._create_tensor(self.loss_input['inputs'])
labels = self._create_tensor(self.loss_input['labels'])
input = inputs[0]
model_output = self.model(input)
loss = Loss()()
loss = loss(model_output, labels)
# Calculate the gradients for the batch of data on this tower.
varlist = [x for x in tf.trainable_variables() if x.name.startswith('tower_%d' % i)]
grads = opt.compute_gradients(loss, varlist)
tower_grads.append(grads)
if i == 0:
self.actor_var = TFVariables(model_output, self.sess)
self.input = input
self.logits = model_output
self.loss = loss
self.inputs.append(inputs)
self.labels.append(labels)
grads = self._average_gradients(tower_grads)
self.train_op = opt.apply_gradients(grads)
self.sess.run(tf.initialize_all_variables())
def _default_model_fn(self, features, labels, mode):
"""Define model_fn used by TensorFlow Estimator.
:params features: input features
:type features: tensorflow tensors
:params labels: label data
:type labels: tensorflow tensors
:params mode: mode of estimator
:type mode: tf.estimator.ModeKeys
:return: tensorflow EstimatorSpec
:rtype: tf.estimator.EstimatorSpec
"""
logging.info('model function action')
self.model.training = mode == tf.estimator.ModeKeys.TRAIN
logits = self.model(features)
logits = tf.cast(logits, tf.float32)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
loss = self.loss(logits, labels)
train_op = None
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.compat.v1.train.get_or_create_global_step()
epoch = tf.cast(global_step, tf.float32) / tf.cast(
len(self.train_loader), tf.float32)
self.optimizer = Optimizer()(distributed=self.distributed)
self.lr_scheduler = LrScheduler()(optimizer=self.optimizer)
self.lr_scheduler.step(epoch)
update_ops = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)
loss_scale = self.config.loss_scale if self.use_amp else 1
minimize_op = self.optimizer.step(loss, loss_scale, global_step)
train_op = tf.group(minimize_op, update_ops)
eval_metric_ops = None
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = self.valid_metrics(logits, labels)
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
def build(self):
"""Build the trainer by assembling the necessary components."""
super().build()
self.optimizer = Optimizer()(model=self.model,
distributed=self.distributed)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
self.lr_scheduler = LrScheduler()(self.optimizer)
# Some trainer has different train batch size from valid batch
self.train_metrics = self._init_metrics()
self.valid_metrics = self._init_metrics()
self._init_horovod_setting()
if self.use_amp:
self.model, self.optimizer = amp.initialize(self.model,
self.optimizer,
opt_level='O1')
def build(self):
"""Build the trainer by assembling the necessary components."""
self._init_hps(self.hps)
logging.debug("Trainer Config: {}".format(self.config))
self.do_validation = self.config.with_valid
self.use_syncbn = self.config.syncbn
if self.use_syncbn and zeus.is_torch_backend():
self.model = apex.parallel.convert_syncbn_model(self.model)
self.train_loader = self._init_dataloader(mode='train')
self.valid_loader = self._init_dataloader(mode='val')
self.batch_num_train = self.train_loader.get_dataset_size() if zeus.is_ms_backend() else len(self.train_loader)
self.batch_num_valid = self.valid_loader.get_dataset_size() if zeus.is_ms_backend() else len(self.valid_loader)
if zeus.is_torch_backend():
self.optimizer = Optimizer()(model=self.model, distributed=self.distributed)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
self.lr_scheduler = LrScheduler()(self.optimizer)
elif zeus.is_ms_backend():
self.optimizer = Optimizer()(model=self.model)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
self.metric_name = self.config.metric().type
# Some trainer has different train batch size from valid batch
self.train_metrics = self._init_metrics() if zeus.is_torch_backend() else None
self.valid_metrics = self._init_metrics()
self._init_horovod_setting()
if self.use_amp and zeus.is_torch_backend():
self.model, self.optimizer = amp.initialize(
self.model, self.optimizer, opt_level='O1')
def _init_train_op(self):
self.inputs = self._create_tensor(self.loss_input['inputs'])
self.labels = self._create_tensor(self.loss_input['labels'])
self.input = self.inputs[0]
logits = self.model(self.input)
self.logits = logits
self.actor_var = TFVariables(logits, self.sess)
loss = Loss()()
self.loss = loss(logits, self.labels)
self.optimizer = Optimizer()(distributed=self.distributed)
grads_and_var = self.optimizer.compute_gradients(self.loss)
grads, var = zip(*grads_and_var)
grads_and_var = list(zip(grads, var))
self.train_op = self.optimizer.apply_gradients(grads_and_var)
self.sess.run(tf.initialize_all_variables())
def model_fn(self, features, labels, mode):
"""Darts model_fn used by TensorFlow Estimator."""
logging.info('Darts model function action')
global_step = tf.compat.v1.train.get_global_step()
train_op = None
if mode == tf.estimator.ModeKeys.TRAIN:
features, valid_features = features['train'], features['valid']
labels, valid_labels = labels['train'], labels['valid']
# update arch
epoch = tf.cast(global_step, tf.float32) / tf.cast(
len(self.trainer.train_loader), tf.float32)
self.trainer.optimizer = Optimizer()(
distributed=self.trainer.distributed)
self.trainer.lr_scheduler = LrScheduler()(self.trainer.optimizer)
self.trainer.lr_scheduler.step(epoch)
update_ops = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)
arch_minimize_op = self.search_alg.step(
valid_x=valid_features,
valid_y=valid_labels,
lr=self.trainer.lr_scheduler.get_lr()[0])
train_op = tf.group(arch_minimize_op, update_ops)
self.model.training = mode == tf.estimator.ModeKeys.TRAIN
logits = self.model(features)
logits = tf.cast(logits, tf.float32)
self.trainer.loss = Loss()()
loss = self.trainer.loss(logits=logits, labels=labels)
if mode == tf.estimator.ModeKeys.TRAIN:
with tf.control_dependencies([train_op]):
weight_ops = self.model.get_weight_ops()
loss_scale = self.trainer.config.loss_scale if self.trainer.use_amp else 1
train_op = self.trainer.optimizer.step(loss, loss_scale,
global_step, weight_ops)
eval_metric_ops = None
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = self.trainer.valid_metrics(logits, labels)
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)